---
title: Riemannian KL Property Analysis
url: https://www.emergentmind.com/topics/riemannian-kurdyka-lojasiewicz-property
type: topic
---

# Riemannian KL Property Analysis

The Riemannian Kurdyka–Łojasiewicz (KL) property extends the key regularity phenomenon underlying sharp global and local convergence of descent-type algorithms from the classical Euclidean to the Riemannian manifold context. It guarantees, for a large class of nonconvex and possibly nonsmooth functions on geodesically complete Riemannian manifolds, a powerful gradient–suboptimality differential inequality near critical points, governing the interplay between residuals and step sizes within iterative optimization schemes. The Riemannian KL property now underpins a host of abstract and algorithmic convergence theorems for inexact proximal, projected, and steepest descent methods on manifolds, and, crucially, provides explicit rates that depend on the KL exponent.  

## 1. Subdifferentials, Gradients, and Quasi-Distances on Riemannian Manifolds

Let $(M,g)$ denote an $n$-dimensional complete Riemannian manifold with metric $g$, distance $d(\cdot,\cdot)$, and geodesic structure. For $f:M\to\mathbb{R}\cup\{+\infty\}$, proper and lower semicontinuous (lsc), the relevant first-order objects are:

- **Fréchet subdifferential** at $x\in\operatorname{dom}f$:
  $$
  \hat\partial f(x)
  =\left\{dh_x: h\in C^1(M),\ f-h \text{ minimal at } x\right\}
  $$
- **Limiting subdifferential**:
  $$
  \partial f(x) = \left\{v\in T_xM: \exists (x_k,v_k)\to(x,v),\ v_k\in\hat\partial f(x_k),\ f(x_k)\to f(x)\right\}
  $$
- **Critical points** are $x$ such that $0\in\partial f(x)$.

A **quasi-distance** $D:M\times M\to\mathbb{R}_+$ satisfies
- $D(x,y)=D(y,x)=0\implies x=y$,
- $D(x,z)\le D(x,y)+D(y,z)$, continuity in the second argument.

The Riemannian gradient $\nabla_M f(x)$ at $x$ satisfies $Df(x)[\xi] = g_x(\nabla_M f(x), \xi)$ for all $\xi\in T_xM$.

## 2. Formal Statement of the Riemannian KL Property

Given a proper lsc $f:M\to\mathbb{R}\cup\{+\infty\}$ and a critical point $\bar x\in\operatorname{dom}\partial f$, $f$ is said to have the Kurdyka–Łojasiewicz property at $\bar x$ if there exist:

- $\eta>0$,
- geodesically convex neighborhood $U$ of $\bar x$,
- $\varphi:[0,\eta)\to\mathbb{R}_+$, continuous, concave, $C^1$ on $(0,\eta)$ with $\varphi(0)=0$ and $\varphi'>0$,

such that for all $x\in U$ with $f(\bar x)<f(x)<f(\bar x)+\eta$,
$$
\varphi'(f(x)-f(\bar x))\cdot \|\nabla_M f(x)\|_x \geq 1
$$
or, in the subdifferential form,
$$
\varphi'(f(x)-f(\bar x))\cdot \operatorname{dist}(0, \partial f(x)) \geq 1.
$$
A global KL function satisfies this property at every point of $\operatorname{dom}\partial f$ [1103.4828], [2502.06183].

## 3. Classes of Functions and Manifold Requirements

The Riemannian KL property is guaranteed for key function classes:

- **Analytic functions** on analytic manifolds (classical Łojasiewicz setup).
- **Definable functions** in an o-minimal structure (semi-algebraic, subanalytic, restricted analytic, etc.).
- **Morse functions** ($C^2$ with nondegenerate critical points): KL with exponent $\tau=1/2$.
- **Nonsmooth, lower-semicontinuous definable functions** on definable submanifolds.

For algorithmic convergence, manifold requirements depend on method:
- Proximal-point: $M$ must be complete (no curvature sign constraint), Hopf–Rinow applies.
- Steepest descent: $M$ assumed Hadamard (complete, simply connected, nonpositive sectional curvature; unique minimizing geodesics and global exponential map) [1103.4828], [2502.06183].

## 4. KL Property, Exponent, and Desingularizing Functions

A typical desingularizing function is $\varphi(s)=c\,s^{1-\tau}$ with $\tau\in[0,1)$, $c>0$. The key inequality becomes:
$$
\|\nabla_M f(x)\|_x \geq c'|f(x)-f(\bar x)|^\tau
$$
for some $c'>0$. The exponent $\theta=\tau$ quantifies critical point sharpness:
- $\theta=0$: “sharp” case—finite-step convergence.
- $\theta\in(0,1/2]$: local linear convergence rate for descent methods.
- $\theta\in(1/2,1)$: local sublinear rate, specifically $\|x_k - \bar x\| = O(k^{-(1-\theta)/(2\theta-1)})$.

For composite or structured potentials (e.g., in ManPPA), KL exponents propagate: if a primal potential $F$ has exponent $\theta$, so does an auxiliary lifted potential $\mathcal{F}$, and conversely under additional regularity [2502.06183].

## 5. KL Property and Convergence of Descent Methods

The KL property is pivotal for establishing full-sequence convergence in inexact (possibly non-Euclidean) descent algorithms, as shown in [1103.4828], [2502.06183]. The abstract convergence framework employs:

- **Proximal-point methods**: For a sequence $(x_k, w_k)$ generated via
  $$
  f(x_{k+1}) + \lambda_k D^2(x_{k+1}, x_k) \leq f(x_k)
  $$
  $$
  w_{k+1}\in \partial f(x_{k+1}),\quad \|w_{k+1}\| \leq b D(x_{k+1}, x_k)
  $$
  boundedness and compactness in $D$ yield convergence of $x_k$ to a critical point.

- **Steepest descent (smooth, $L$-Lipschitz gradient)**: For $f\in C^1(M)$,
  $$
  f(x_{k+1}) + r_1 D^2(x_{k+1}, x_k) \le f(x_k)
  $$
  $$
  \|\nabla f(x_k)\| \le r_2 D(x_{k+1}, x_k)
  $$
  With $D=d$ (the Riemannian distance), on Hadamard manifolds, this applies to exponential-mapped iterates with Armijo or fixed-step rules.

Fundamental technical estimates, such as
$$
\varphi(f(x_k)-f(\bar x)) - \varphi(f(x_{k+1})-f(\bar x)) \ge \frac{D^2(x_{k+1}, x_k)}{b D(x_k, x_{k-1})}
$$
ensure iterates remain in a KL ball and the steps sum to a finite total, so the full sequence converges [1103.4828].

## 6. Application: Manifold Proximal Point and Related Algorithms

For constrained and structured problems, the Riemannian KL property enables convergence analysis of manifold-based algorithms such as the Manifold Proximal Point Algorithm (ManPPA). For instance, on $M=S^{n-1}$ (unit sphere),
$$
\min_{x\in M} f(x) = \sum_{i=1}^p \phi(|y_i^T x|)
$$
with $\phi$ a concave regularizer, KL analysis applies both to the primal objective $F(x)=f(x)+I_M(x)$ and to lifted auxiliary potentials $\mathcal{F}(x,u)$. Concrete residual-step bounds and descent inequalities facilitate direct application of the abstract KL convergence theorems, guaranteeing:

- Global convergence of the sequence to a critical point.
- Explicit linear or sublinear rates determined by the KL exponent $\theta$.
- Finite-step convergence when $\theta=0$, such as in the weak-sharp minima setting.

The propagation and equivalence of KL exponents between potential functions are established for key problem classes, making rate conclusions robust to various reformulations [2502.06183].

## 7. Examples and Scope of KL Functions on Manifolds

The class of KL functions on manifolds is broad, encompassing:

- **Real-analytic functions** on analytic manifolds (guaranteed classical Łojasiewicz property).
- **Definable functions** (semi-algebraic, subanalytic, restricted analytic) on definable submanifolds, ensuring the KL property via o-minimal structure theory.
- **Morse functions:** Smooth objectives with nondegenerate critical points, with exponent $\theta=1/2$.
- **Nonsmooth, prox-regular, semialgebraic functions**—ubiquitous in composite optimization, statistical estimation, and signal processing on manifolds [1103.4828], [2502.06183].

The Riemannian KL property thereby serves as the geometric-analytic backbone for convergence analysis, rate guarantees, and regularity theory of nonconvex optimization across a broad swath of smooth and nondifferentiable problems on manifolds.

Source: https://www.emergentmind.com/topics/riemannian-kurdyka-lojasiewicz-property